Where Should AI Sit in the Board Decision-Making Process?

A board pack arrives with a market analysis a machine drafted in under a minute, a risk model built from scenarios nobody in the room chose, and a recommendation with a confidence score attached to it. Nobody approved any of this in principle. It simply started happening, one meeting at a time, until AI was doing real work inside the board's decisions without anyone deciding where that work should stop. 

Sixty per cent of executives now regularly use AI to support their decisions, according to Deloitte's 2026 Global Human Capital Trends research, and Gartner projects that by 2027, half of all business decisions will be augmented or automated by AI agents. Boards are catching up to that reality rather than setting the pace of it. Most of the anxiety around AI governance isn't really about whether to use the technology. It's about the fact that AI has already found its way into specific moments of the decision, preparation, analysis, the recommendation itself, without a board ever sitting down and deciding that was the right place for it. 

Paul Armstrong, founder of TBD Group, framed the underlying problem well in Clyde & Co's Risk Quarterly: AI systems carry different levels of delegated authority, from assistance to advice to full autonomy, and collapsing that distinction obscures “the point at which responsibility transfers from human judgement to system behaviour.” A board that hasn't located that point, deliberately, at every stage of how it decides things, has effectively let AI find its own level instead. 

In this article: 

• The four places AI does its work 
• Where the line has to hold 
• How better board infrastructure can help 

The four places AI does its work 

In our experience,  it’s often helpful to consider board decision across four stages: preparation, scenario analysis, deliberation and follow-up. AI's usefulness, and its risk, change sharply depending on which one it's touching. 

Preparation is the easy case. Before a meeting, AI can synthesise long or technical material and flag what needs a director's attention, which mostly just saves time that used to go into reading rather than thinking. The one condition that matters here is timing: material has to reach directors early enough, and securely enough, that they can form their own view of it, not receive a summary for the first time in the room. 

Scenario analysis is where AI earns something closer to real credit. A model can generate far more variations of a future than a workshop ever could, testing assumptions about rates, competitors, regulation and demand in combinations a small team wouldn't have time to work through by hand. The job is to expand what the board argues about. It's not to hand back a single ranked answer dressed up as analysis, since that ranking reflects the model's assumptions about risk, not the board's. 

Then there's deliberation, and this is where the caution has to be sharpest, because an AI-generated recommendation tends to arrive looking finished right at the moment a board most needs to pick it apart. As one Liberty Mutual leader put it to MIT Sloan Management Review, once a model has spoken the real question becomes “who gets to disagree with it, and how fast?”  

That's the part no system can do on a board's behalf. It takes someone in the room who sees the problem differently, who pushes on the one assumption nobody else questioned, who is willing to disagree with a conclusion that reads cleanly on the page.  As Julie Baddeley notes in a recent episode of the Boardroom Confidential podcast, “"It's very easy for a board to slip into groupthink where everybody's reinforcing everybody else's thinking. If you're an outlier, it's not a very comfortable place to be."  

Treat the recommendation as the opening move in an argument and AI has done its job. Treat it as the answer and the board has quietly handed away the one part of the work that was never AI's to do. 

Follow-up gets less attention than the other three, which is a mistake, because it's where good deliberation most often evaporates. AI can turn a decision into a proper record: the reasoning, the actions, who owns what and by when, so a strong discussion doesn't just live in the memories of whoever was in the room. Some of that needs to happen off the standing calendar entirely. A board that can't securely review, discuss and, where it's warranted, approve something between scheduled meetings is stuck waiting for the next date on the calendar to catch up with a decision that already needed making. 

Where the line has to hold 

None of the above changes who is accountable. Whatever AI contributed at any stage, the board carries the decision, in full. 

In practice that means a few things stay stubbornly human. Someone has to work out what an AI-generated insight means for this specific company rather than taking its conclusion at face value. Someone has to press on the assumption nobody bothered to say out loud. And somebody, eventually, has to decide whether the evidence in front of the board is enough to act on, which is a judgement call no model gets handed. 

The risk here isn't abstract. Deloitte's 2026 research on AI and decision-making found that “black box” algorithms make it hard to trace who's responsible for what, that people report feeling less ownership over decisions AI helped shape, and, more unsettling, that people appear to grow more willing to cut corners once a decision has been delegated to a machine. None of that is an argument against using AI here. It's an argument for being precise about exactly which part of the decision it's allowed to touch. 

Security is a good place to see this precision in action. AI can behave like something close to a virtual CISO inside the process, catching a data-exposure risk buried in a vendor contract or an integration plan long before it would otherwise surface. That's real, useful work. It's also exactly the kind of finding that belongs in front of an actual CISO, or an equivalent specialist, before the board commits to anything built on top of it. AI surfacing the risk is preparation. A qualified person weighing what it means is judgement, and that handoff should happen every time, not when someone remembers to ask for it. 

A board that has placed AI on purpose, deliberately, can usually answer a few things without much hesitation: which stage AI is and isn't currently touching, whether a director has ever pushed back on an AI-generated recommendation and won the argument, and whether a security concern AI first flags reliably reaches a specialist rather than getting absorbed into a tidy summary. A board that has to think hard about any of these has let AI settle into the process by drift rather than by decision. 

How better board infrastructure can help 

None of this works if the tool doing the supporting lives outside the board's secure workflow, in some separate app nobody's tracking against the record. 

Purpose-built meeting solutions such as Sherpany offer bespoke AI tooling to support the challenges of modern governance. Sherpany's AI features sit inside the workflow boards follow, rather than beside it. Before a meeting, they help directors work through board papers on their own, surfacing what needs closer attention, and summarising meeting materials, so preparation increases in efficiency and is focused on materials that are properly permissioned and secure. During discussion, the same system keeps challenges and questions attached to what they're about. Afterward, it captures decisions and follow-up actions properly, so none of it depends on someone's inbox or a set of personal notes nobody else can check. Sherpany’s AI meeting minutes ensure that the rationale behind board decisions, which is more critical during the age of AI than ever, is never lost. 

None of that decides anything for the board. It just means the deciding happens somewhere documented, secure, and built to hold up later, rather than scattered across whatever tool happened to be open at the time. For a fuller framework on building the governance habits behind this, our guide, 5 Practical Ways to Build an AI-Ready Board, goes further into how that oversight gets structured. 

AI clearly belongs in how boards decide things now, and mostly by default rather than by design so far. The narrower question is the one worth answering: where does it earn its keep, and where does the board's own judgement have to do the work nothing else can. Get that placement right stage by stage, and AI sharpens preparation and widens analysis without ever touching the part of the job that was always the board's to carry. 

If you would like to see how Sherpany supports AI-enabled governance across the board decision lifecycle, book a free demo and find out how Sherpany can help.